Abstract
Deep learning-based ischemic stroke lesion segmentation from non-contrast computed tomography (NCCT) is constrained by the limited availability of voxel-level annotations, especially in low-resource clinical settings. Two commonly explored approaches for improving data efficiency are self-supervised learning (SSL), which learns transferable representations from unlabeled data, and data augmentation, which increases the diversity of available labeled examples. However, their relative contribution under extreme annotation scarcity remains unclear. In this work, we perform a controlled evaluation of SSL pretraining and segmentation-aware augmentation for acute ischemic stroke lesion segmentation using a locally acquired NCCT dataset. The dataset comprises 29 manually annotated and 178 unlabeled CT volumes collected from a tertiary hospital. We adapt a segmentation-aware augmentation framework to three-dimensional stroke CT by incorporating lesion-localized elastic deformation together with CT-compatible geometric and intensity transformations. Using Swin UNETR as the segmentation architecture, we evaluate four training configurations by independently varying the use of SSL pretraining and segmentation-aware augmentation. Five-fold cross-validation shows that segmentation-aware augmentation provides the largest performance improvement, increasing the mean Dice score from 0.502 to 0.555 without SSL pretraining and from 0.511 to 0.551 with SSL pretraining. In comparison, SSL pretraining alone provides only marginal improvement when using the original annotated dataset and does not yield additional gains when combined with augmentation. These findings indicate that segmentation-aware augmentation can substantially improve data efficiency in severely annotation-limited stroke CT segmentation, while the benefit of SSL may depend on the scale and diversity of the available unlabeled data.
Links to Paper and Supplementary Materials
Main Paper (Open Access Version): https://papers.miccai.org/miccai-2026-sat/paper/AFRICAI_044.pdf
SharedIt Link: Not yet available
SpringerLink (DOI): Not yet available
Supplementary Material: Not Submitted
Link to Open Review
Open Review Page: https://openreview.net/forum?id=b488me5OKD
BibTex
@InProceedings{AyiLiv_Selfsupervised_MICCAISAT2026,
author = { Ayivor, Livingstone Eli AND Adjei, Prince Ebenezer AND Akpaloo, Belinda Esinam AND Otoo, Kojo Obed AND Nti Koduah, Joseph D. A. AND Amuasi, John},
title = { { Self-supervised Learning Versus Segmentation-aware Augmentation for Acute Ischemic Stroke CT Segmentation Under Annotation Scarcity } },
booktitle = {Medical Image Computing and Computer Assisted Intervention -- MICCAI 2026 Workshops and Challenges},
year = {2026},
publisher = {Springer Nature Switzerland},
volume = {LNCS 17264},
month = {pending},
page = {pending}
}